search

s-gupta / rcnn-depth

Learning Rich Features from RGB-D Images for Object Detection and Segmentation
BSD 2-Clause "Simplified" License
365 stars 150 forks source link

Reproducing results in the paper #1

Closed varun-nagaraja closed 10 years ago

varun-nagaraja commented 10 years ago

Is there a script to run object detection on all the images and evaluate with the ground truth? It would be great if you could tell me any other parameters used (like thresholds) to obtain the results in the paper.

s-gupta commented 10 years ago

Yes, there is. You will have to do the following.

  1. Compute UCMs by: jobName = 'edges_to_ucms'; script_regions;
  2. Compute MCG regions and boxes: jobName = 'region-detect'; script_regions;
  3. Save color and hha images by:jobName = 'save_hha'; script_detection; jobName = 'save_color'; script_detection;
  4. Compute CNN features: jobName = 'hha_cache_features'; script_detection; jobName = 'color_cache_features'; script_detection;
  5. Train and test RCNN SVMs jobName = 'rcnn_train'; script_detection; (It will use the pretrained model by default if you downloaded the pre-trained models).

You will have to modify jobParallel, to run the code on your cluster / machine (It essentially runs the function handle on the arguments). I am sorry this is a bit of work. If you just want to play around with the object detectors, I can upload the precomputed UCMs, regions and boxes (it is quite some data, and I am investigating what is the best way to make it available). You will still have to execute steps 3 to 5, though.

Let me know if you have any follow-up questions.

varun-nagaraja commented 10 years ago

I should be able to run through all the 5 steps. Thanks for this guide and in general, for this code. This is very helpful for me.